Re: Instability in select_parallel regression test

On Mon, Feb 20, 2017 at 7:52 AM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
> On Sun, Feb 19, 2017 at 8:32 PM, Robert Haas <robertmhaas(at)gmail(dot)com> wrote:
>> On Sun, Feb 19, 2017 at 6:50 PM, Amit Kapila <amit(dot)kapila16(at)gmail(dot)com> wrote:
>>> To close the remaining gap, don't you think we can check slot->in_use
>>> flag when generation number for handle and slot are same.
>>
>> That doesn't completely fix it either, because
>> ForgetBackgroundWorker() also does
>> BackgroundWorkerData->parallel_terminate_count++, which we might also
>> fail to see, which would cause RegisterDynamicBackgroundWorker() to
>> bail out early. There are CPU ordering effects to think about here,
>> not just the order in which the operations are actually performed.
>>
>
> Sure, I think we can attempt to fix that as well by adding write
> memory barrier in ForgetBackgroundWorker().

I don't think so.

> The main point is if we
> keep any loose end in this area, then there is a chance that the
> regression test select_parallel can still fail, if not now, then in
> future. Another way could be that we can try to minimize the race
> condition here and then adjust the select_parallel as suggested above
> so that we don't see this failure.

My guess is that if we apply the fix I suggested above, it'll be good
enough. If that turns out not to be true, then I guess we'll have to
deal with that, but why not do the easy thing first?

--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company